When I calculate KDE for points distribution the greatest difficulty is to define the best bandwidth. User-friendly tools that define the best bandwidth could be very useful for creating more accurate KDE surface.
If you still have ArcGIS version 9, you could install the Home Range Tools Add In, and with this you can calculate kernel densities and select different bandwidth choices. You can use the function "proportion of reference bandwidth" to calculate the ad hoc bandwidth where the outer isoline is just still continuous and not yet broken up. This is then the best bandwidth for your dataset. If you are interested I could search for the papers again where this method is described. I already attach you the HRT manual if you want to have a look. But HRT does not work with ArcGIS 10.
You can use either the new Home Range Version 3 plugin for Quantum GIS (http://hub.qgis.org/projects/quantum-gis/wiki/HomeRange_plugin) or (IMHO better) the R package adehabitatHR (http://cran.r-project.org/web/packages/adehabitatHR/).
More in detail, finding "optimal" values for the "infamous" h parameter is still a topic under discussion. My suggestion is a tad biased, but you can look at a method that me and my colleagues worked out (see link below). The method is implemented for "old" arehabitat R package and I'm working on a new implementation for adehabitatHR. Code is available as R package on RForge (https://r-forge.r-project.org/projects/hrtools/): it takes just _three rows_ of R code. :)
Article Radio-tracking squirrels: Performance of home range density ...
You can use the tool Hawth's Analysis Tools for ArcGIS (http://www.spatialecology.com/htools/kde.php), or use the QGIS plugin called Home-range analyzes (http://hub.qgis.org/projects / quantum-gis/wiki/HomeRange_plugin), but as mentioned in other reviews, you'll have to perform several simulations to see which is the bandwidth value that best fits your data.
Add-on: usually, the "best" bandwidth value is found via least-squares cross-validation. adrhabitat/arehabitatHR kernelUD function finds that (h="LSCV" option ). Sometimes, the minimization routine used fails to converge in searching for a global minimum. AFAIK, this is a well-known issue and no final solution has been found yet.
The current iteration of Hawth Tools, the stand-alone Geospatial Modeling Environment (GME), contains a Kernel Density function. This includes the bandwidth estimators: SCV, BCV, BCV2, PLUGIN, LSCV, CVh. http://www.spatialecology.com/gme/kde.htm. GME can read/write features and rasters in a Geodatabase.